September 28, 2026

Expertise and Trust in Humanities AI

What AI changes about humanities research, and how it recalibrates trustwork.

Fred Gibbs: fredgibbs.net
History: history.unm.edu
Amaranth: amaranth.unm.edu

Provocation

We can already trust AI.

Do you trust me? That is the real issue. AI is trustworthy in the sense that skilled users know what to expect and that AI output is already consistent (while improving). The question we need to answer is how do we know to trust those who use it?

Two printings of the same lunar image: a fine etching and a crude woodcut copy

Venice 1610 · etched Frankfurt 1610 · pirated

Precedent 1

Print was not automatically trustworthy.

The same Moon, the same year. Galileo’s own etching on the left; on the right a Frankfurt piracy, recut in wood and printed upside down with stronger contrast. The latter was far more reprinted. Which is, or becomes, true?

1:20–2:00

Before any of my own work, two slides of history — because that question is old, and we are not the first people to face it.

Print did not arrive trustworthy. More copies meant more chances for corruption, and no reader could check a text against an exemplar they would never see. What eventually made print credible was not the press; it was a century of institution-building around it — correctors on the payroll, printing privileges, colophons naming who was answerable, a trade register. Adrian Johns’s argument: fixity is not a property of print. It is something people had to manufacture, and keep manufacturing.

Now the picture, which is the argument in one object. Galileo published in Venice in March 1610 with copper etchings made from his own wash drawings. Within months a Frankfurt printer reissued it without permission, and in the hurry the plates were recut as woodcuts — cheaper, coarser, and set into the forme upside down. Point at the big crater: low on the terminator on the left, high on the right.

Then the sting, which is the reason this slide is here. The pirated woodcuts were the ones later editions copied, and the ones that went into moon handbooks for centuries. Scholars who never saw a first edition concluded that Galileo was a crude draughtsman. The bad copy outcompeted the good one, and the author took the reputational damage for it.

Say the parallel lightly. The next slide completes it.

Sources: Adrian Johns, The Nature of the Book (1998). Edition history and the upside-down woodcuts: Linda Hall Library, “The Face of the Moon,” section B (1610–1700). Images: Sidereus Nuncius, Venice: Baglioni, 1610 (Smithsonian Libraries copy, Internet Archive Sidereusnuncius00Gali) and Frankfurt: Palthenius, 1610 (Boston Public Library copy, Internet Archive sidereusnunciusm00gali_0) — both public domain.

Precedent 2

Science had to build its trust network.

Joseph Wright of Derby, 1768 — a century after Boyle’s pump, and a fact is still made by gathering people into a room and letting them watch the bird die. The Royal Society, the witnessed demonstration, reports written so a distant reader could witness at second hand: trust infrastructure, and it took generations to build. Does AI threaten that?

2:00–2:40

Same problem, next century, and this one is a laboratory.

The experimental philosophy produced claims almost nobody could check: what happened inside Boyle’s air pump was believable only if you trusted the people standing there. So the period built machinery for exactly that — the Royal Society as a standing body of credible witnesses, demonstrations performed in front of an audience rather than reported afterwards, and papers written in enough circumstantial detail that a reader in another country could witness them at second hand. Shapin and Schaffer call that virtual witnessing.

Now the date on this picture. It is 1768 — more than a hundred years after Boyle. They are still doing it: still gathering people into a room, still watching each other’s faces. That is the part I want to land. The network that makes knowledge trustworthy was not switched on. It was built slowly, argued over, and it took generations.

We are about two years into ours.

Sources: Steven Shapin and Simon Schaffer, Leviathan and the Air-Pump (1985); Steven Shapin, A Social History of Truth (1994). Image: Joseph Wright of Derby, An Experiment on a Bird in an Air Pump, 1768, National Gallery, London — public domain.

Amaranth, UNM's digital humanities studio

amaranth.unm.edu

A studio, not a service desk.

Amaranth is UNM’s digital humanities studio: a research and teaching space where faculty, students, and community partners investigate how digital methods can deepen humanistic inquiry, strengthen public scholarship, and build the digital literacy students need for an AI-shaped world.

AI Sketchbook

Field notes, especially the failures.

A shared space where colleagues try something, write down what happened, and tag it. Fifteen sketches so far, each marked tested or untested — because the honest answer is usually that it half worked.

Does AI encourage local knowledge more valuable?

Humanities workflows

Scale is no longer the obstacle.

Reading across thousands of pages, drafting transcripts for a whole interview collection, describing a few hundred photographs so they can be searched — all of it is now a few weeks of supervised work rather than a years-long funded project. The main constraint is no longer capacity, it is judgment and competency.

4:15–5:00

The first focus: bringing AI into humanities research itself, not just into the classroom.

Be concrete about what changed. It isn’t that the machine is smart. It’s that three or four specific bottlenecks — transcription, description, normalization of messy archival data, first-pass reading at scale — got cheap enough that a single researcher or a class can now attempt a collection that used to need a grant and a team.

And be equally concrete about what didn’t change: every one of those outputs is wrong somewhere, usually in the places that carry the most meaning. So the work didn’t disappear; it moved to checking, and checking requires more expertise than producing did.

Next on this list is mapping: georectification, layer alignment and feature extraction are exactly the technical drudgery that keeps spatial analysis out of a humanities course. AI does the lift; the reading stays ours.

AI for humanities research

New reach, and new ways to be wrong.

01 / ACROSS TEXTS

Read ten thousand pages you haven’t read.

Themes traced across a whole corpus, language shifting over decades, patterns no close reading would surface.

You are now vouching for a pattern you did not see.

02 / IN SPEECH

Forty hours of tape, searchable tomorrow.

Draft transcripts of an entire oral history collection in days rather than months, and build an index.

It mangles the names, the Spanish, and the place words — which is most of what the collection is about.

03 / IN THE LITERATURE

Transcend disciplines.

The major positions, the live debates, the search terms, chart a route into a literature you have not read.

It invents sources. Or does it?

5:00–5:50

Three real workflows, and I want the pairs rather than the promises.

The top line of each is a genuine change in reach. Work that needed a team and a grant is now a semester project — that part is not hype.

The bottom line is what it costs, and notice the pattern: in every case the machine is weakest exactly where the material is most specific. It reads a corpus but cannot tell you which pattern matters. It transcribes fluently and then flattens the proper nouns, the Spanish, and the acequia vocabulary — the words that carry the local knowledge. It maps a literature and salts the map with books that do not exist.

So the competence required went up, not down. You now need enough expertise to catch a confident machine in the places it is confidently wrong — and that is precisely the expertise a reader outside the field cannot check.

Which is the trust problem in its working form: not “is the AI reliable,” but “how would you know whether I checked.”

Source: amaranth.unm.edu/projects/ai-humanities — the workflows and their stated caveats.

Reconstruction

Every reconstruction is an argument with the confidence turned up.

LiDAR and photogrammetry fail visibly: a gap in a point cloud looks like a gap. A generated elevation looks finished whether or not anyone alive knows what was there. If we build these, the drawing convention has to carry the evidence — which is a discipline architectural illustration has had for a century and the new tools do not.

AI-generated image

Ninety-four per cent confident. Of what?

Both halves are generated — the “reality” on the left is not a photograph either. Nothing in the reconstruction is sourced to anything, and the number measures nothing at all.

6:35–7:30

Let them look before you say anything. Someone will read the panel out loud.

What the picture promises is a clean split: measured fact on the left, reconstruction on the right, and a seam down the middle you can see. What it actually delivers is a single generated image in which both halves are invented, including the ruin. There is no survey behind the wall on the left and no source behind the roof beams, the plaster, the hanging corn, or the people on the right.

Then the number. Ninety-four per cent confident — of what, measured against what? A confidence score needs a ground truth to be a percentage of, and for a building nobody alive has stood inside there isn’t one. It is a number with the grammar of evidence and none of the substance, and it is the most persuasive thing on the screen. If you remember one image from this talk, make it this one.

The remedy is old and this field already owns it. Reconstruction drawings have distinguished extant fabric from conjecture for a century, with line weight and hatching. I have a drawn version of this for the Alvarado Hotel — solid for what is photographed, dashed for what is inferred from building type, dotted for what I would be inventing. It is less impressive and it is honest, and the discipline has to be in the drawing convention rather than in a disclaimer nobody reads.

Sources: image generated with Google Gemini for this talk, September 2026. The panel’s own claims — Aztec Ruins National Monument, Pueblo III — are the model’s, not mine; do not repeat them as fact.

Teaching

Teach peer review

When public scholarship is the default output, the possibilities and incentives change from day one: students write for readers, cite for strangers, and think about trust.

UNM Campus Histories — a multi-semester project. Each cohort builds on what the last one made.

Oral Histories of Farming along the Middle Rio Grande

Community storytelling

The knowledge is already in the community.

Farmers along the Middle Rio Grande hold knowledge about water, soil, and land use that exists in no document anywhere. Documented by students in Geography and Environmental Studies and Sustainability Studies.

Is this a conversation or an archive?

Farmer profiles, each crediting the narrator, the interviewer and the authors

Stewardship

Do human voice matter more?

Every profile credits the farmer, the student who ran the interview, and the students who wrote it up, the instructor who guided the class.

Can AI tools cite this?

Five 3D-printed medieval capitals in different filaments, laid out on a white table

Amaranth · 3D pedagogy

Scan it, print it, hold it.

We printed five medieval capitals from Scan the World. Set them on a table and the Romanesque-to-Gothic transition stops being a claim in a textbook: it becomes something you see by turning two objects.

Amaranth · 3D pedagogy

Then tell people what to look at.

The poster is the contextualising layer — the digitizing workflow, the feature-by-feature comparison, and the museum that holds each original.

Speculative fiction

Ask it for a place that never existed.

The most useful exercise I have found puts all of this together: have a model generate a landscape, a building, a settlement — something plausible and wholly invented — and then argue about it. Practice the methods of AI critique.

AI-generated image

What here is real? What isn’t?

A generated model landscape of cliff dwellings. No site, no survey, no scale. Students are asked which parts they would defend, on what evidence, and what would settle it.

12:30–13:40

Put it up and let the room work on it for a moment before you say anything. Say clearly that it is AI-generated — after the last two slides it would be absurd not to.

Then the questions, in order. What is doing the work of realism here? The contour terrain reads as a survey model, which is a claim about measurement that nothing in this image is entitled to make. Where are the dwellings placed, and is that where they are actually built? What is the construction — is that masonry coursing or is it texture? What is missing that would be there: middens, water, paths, agriculture, any sign of how anyone got in or out?

A student who knows the Southwest will get to “the siting is wrong” fast. Getting them to say how they know is the entire pedagogical payload.

The questions

These slides are AI-generated

But what does that mean?

Does AI use matter when carefully mediated?

Does AI use matter, or AI credentials?

Fred Gibbs · fredgibbs.net · amaranth.unm.edu